[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M730ZM322ZW6SC1JTGAC61Z8":3,"skill-same:01M730ZM322ZW6SC1JTGAC61Z8":49},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":26,"commitSha":27,"skillDirectory":28,"forksCount":29,"repositoryHtmlUrl":30,"repositoryHomepageUrl":6,"ownerHtmlUrl":31,"ownerBlogUrl":32,"ownerDescription":33,"updatedAt":34,"install":35,"resources":48},"01M730ZM322ZW6SC1JTGAC61Z8","gan-style-harness",null,"GAN-inspired Generator-Evaluator agent harness for building high-quality applications autonomously. Based on Anthropic's March 2026 harness design paper.","SriramGandhiS\u002FSkills","SriramGandhiS","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F211849986?v=4","01M72WQCW12BYX71C6WN584TTA","Sriram S","01M72WQCWTJ2RD0ET1X31AF1RC",1,"MIT","https:\u002F\u002Fgithub.com\u002FSriramGandhiS\u002FSkills\u002Fblob\u002Fmaster\u002Fskills\u002Fgan-style-harness\u002FSKILL.md",[],[],"2026-11-11 15:11:14",false,"---\nname: gan-style-harness\ndescription: \"GAN-inspired Generator-Evaluator agent harness for building high-quality applications autonomously. Based on Anthropic's March 2026 harness design paper.\"\norigin: ECC-community\ntools: Read, Write, Edit, Bash, Grep, Glob, Task\n---\n\n# GAN-Style Harness Skill\n\n> Inspired by [Anthropic's Harness Design for Long-Running Application Development](https:\u002F\u002Fwww.anthropic.com\u002Fengineering\u002Fharness-design-long-running-apps) (March 24, 2026)\n\nA multi-agent harness that separates **generation** from **evaluation**, creating an adversarial feedback loop that drives quality far beyond what a single agent can achieve.\n\n## Core Insight\n\n> When asked to evaluate their own work, agents are pathological optimists — they praise mediocre output and talk themselves out of legitimate issues. But engineering a **separate evaluator** to be ruthlessly strict is far more tractable than teaching a generator to self-critique.\n\nThis is the same dynamic as GANs (Generative Adversarial Networks): the Generator produces, the Evaluator critiques, and that feedback drives the next iteration.\n\n## When to Use\n\n- Building complete applications from a one-line prompt\n- Frontend design tasks requiring high visual quality\n- Full-stack projects that need working features, not just code\n- Any task where \"AI slop\" aesthetics are unacceptable\n- Projects where you want to invest $50-200 for production-quality output\n\n## When NOT to Use\n\n- Quick single-file fixes (use standard `claude -p`)\n- Tasks with tight budget constraints (\u003C$10)\n- Simple refactoring (use de-sloppify pattern instead)\n- Tasks that are already well-specified with tests (use TDD workflow)\n\n## Architecture\n\n```\n                    ┌─────────────┐\n                    │   PLANNER   │\n                    │  (Opus 4.6) │\n                    └──────┬──────┘\n                           │ Product Spec\n                           │ (features, sprints, design direction)\n                           ▼\n              ┌────────────────────────┐\n              │                        │\n              │   GENERATOR-EVALUATOR  │\n              │      FEEDBACK LOOP     │\n              │                        │\n              │  ┌──────────┐          │\n              │  │GENERATOR │--build-->│──┐\n              │  │(Opus 4.6)│          │  │\n              │  └────▲─────┘          │  │\n              │       │                │  │ live app\n              │    feedback             │  │\n              │       │                │  │\n              │  ┌────┴─────┐          │  │\n              │  │EVALUATOR │\u003C-test----│──┘\n              │  │(Opus 4.6)│          │\n              │  │+Playwright│         │\n              │  └──────────┘          │\n              │                        │\n              │   5-15 iterations      │\n              └────────────────────────┘\n```\n\n## The Three Agents\n\n### 1. Planner Agent\n\n**Role:** Product manager — expands a brief prompt into a full product specification.\n\n**Key behaviors:**\n- Takes a one-line prompt and produces a 16-feature, multi-sprint specification\n- Defines user stories, technical requirements, and visual design direction\n- Is deliberately **ambitious** — conservative planning leads to underwhelming results\n- Produces evaluation criteria that the Evaluator will use later\n\n**Model:** Opus 4.6 (needs deep reasoning for spec expansion)\n\n### 2. Generator Agent\n\n**Role:** Developer — implements features according to the spec.\n\n**Key behaviors:**\n- Works in structured sprints (or continuous mode with newer models)\n- Negotiates a \"sprint contract\" with the Evaluator before writing code\n- Uses full-stack tooling: React, FastAPI\u002FExpress, databases, CSS\n- Manages git for version control between iterations\n- Reads Evaluator feedback and incorporates it in next iteration\n\n**Model:** Opus 4.6 (needs strong coding capability)\n\n### 3. Evaluator Agent\n\n**Role:** QA engineer — tests the live running application, not just code.\n\n**Key behaviors:**\n- Uses **Playwright MCP** to interact with the live application\n- Clicks through features, fills forms, tests API endpoints\n- Scores against four criteria (configurable):\n  1. **Design Quality** — Does it feel like a coherent whole?\n  2. **Originality** — Custom decisions vs. template\u002FAI patterns?\n  3. **Craft** — Typography, spacing, animations, micro-interactions?\n  4. **Functionality** — Do all features actually work?\n- Returns structured feedback with scores and specific issues\n- Is engineered to be **ruthlessly strict** — never praises mediocre work\n\n**Model:** Opus 4.6 (needs strong judgment + tool use)\n\n## Evaluation Criteria\n\nThe default four criteria, each scored 1-10:\n\n```markdown\n## Evaluation Rubric\n\n### Design Quality (weight: 0.3)\n- 1-3: Generic, template-like, \"AI slop\" aesthetics\n- 4-6: Competent but unremarkable, follows conventions\n- 7-8: Distinctive, cohesive visual identity\n- 9-10: Could pass for a professional designer's work\n\n### Originality (weight: 0.2)\n- 1-3: Default colors, stock layouts, no personality\n- 4-6: Some custom choices, mostly standard patterns\n- 7-8: Clear creative vision, unique approach\n- 9-10: Surprising, delightful, genuinely novel\n\n### Craft (weight: 0.3)\n- 1-3: Broken layouts, missing states, no animations\n- 4-6: Works but feels rough, inconsistent spacing\n- 7-8: Polished, smooth transitions, responsive\n- 9-10: Pixel-perfect, delightful micro-interactions\n\n### Functionality (weight: 0.2)\n- 1-3: Core features broken or missing\n- 4-6: Happy path works, edge cases fail\n- 7-8: All features work, good error handling\n- 9-10: Bulletproof, handles every edge case\n```\n\n### Scoring\n\n- **Weighted score** = sum of (criterion_score * weight)\n- **Pass threshold** = 7.0 (configurable)\n- **Max iterations** = 15 (configurable, typically 5-15 sufficient)\n\n## Usage\n\n### Via Command\n\n```bash\n# Full three-agent harness\n\u002Fproject:gan-build \"Build a project management app with Kanban boards, team collaboration, and dark mode\"\n\n# With custom config\n\u002Fproject:gan-build \"Build a recipe sharing platform\" --max-iterations 10 --pass-threshold 7.5\n\n# Frontend design mode (generator + evaluator only, no planner)\n\u002Fproject:gan-design \"Create a landing page for a crypto portfolio tracker\"\n```\n\n### Via Shell Script\n\n```bash\n# Basic usage\n.\u002Fscripts\u002Fgan-harness.sh \"Build a music streaming dashboard\"\n\n# With options\nGAN_MAX_ITERATIONS=10 \\\nGAN_PASS_THRESHOLD=7.5 \\\nGAN_EVAL_CRITERIA=\"functionality,performance,security\" \\\n.\u002Fscripts\u002Fgan-harness.sh \"Build a REST API for task management\"\n```\n\n### Via Claude Code (Manual)\n\n```bash\n# Step 1: Plan\nclaude -p --model opus \"You are a Product Planner. Read PLANNER_PROMPT.md. Expand this brief into a full product spec: 'Build a Kanban board app'. Write spec to spec.md\"\n\n# Step 2: Generate (iteration 1)\nclaude -p --model opus \"You are a Generator. Read spec.md. Implement Sprint 1. Start the dev server on port 3000.\"\n\n# Step 3: Evaluate (iteration 1)\nclaude -p --model opus --allowedTools \"Read,Bash,mcp__playwright__*\" \"You are an Evaluator. Read EVALUATOR_PROMPT.md. Test the live app at http:\u002F\u002Flocalhost:3000. Score against the rubric. Write feedback to feedback-001.md\"\n\n# Step 4: Generate (iteration 2 — reads feedback)\nclaude -p --model opus \"You are a Generator. Read spec.md and feedback-001.md. Address all issues. Improve the scores.\"\n\n# Repeat steps 3-4 until pass threshold met\n```\n\n## Evolution Across Model Capabilities\n\nThe harness should simplify as models improve. Following Anthropic's evolution:\n\n### Stage 1 — Weaker Models (Sonnet-class)\n- Full sprint decomposition required\n- Context resets between sprints (avoid context anxiety)\n- 2-agent minimum: Initializer + Coding Agent\n- Heavy scaffolding compensates for model limitations\n\n### Stage 2 — Capable Models (Opus 4.5-class)\n- Full 3-agent harness: Planner + Generator + Evaluator\n- Sprint contracts before each implementation phase\n- 10-sprint decomposition for complex apps\n- Context resets still useful but less critical\n\n### Stage 3 — Frontier Models (Opus 4.6-class)\n- Simplified harness: single planning pass, continuous generation\n- Evaluation reduced to single end-pass (model is smarter)\n- No sprint structure needed\n- Automatic compaction handles context growth\n\n> **Key principle:** Every harness component encodes an assumption about what the model can't do alone. When models improve, re-test those assumptions. Strip away what's no longer needed.\n\n## Configuration\n\n### Environment Variables\n\n| Variable | Default | Description |\n|----------|---------|-------------|\n| `GAN_MAX_ITERATIONS` | `15` | Maximum generator-evaluator cycles |\n| `GAN_PASS_THRESHOLD` | `7.0` | Weighted score to pass (1-10) |\n| `GAN_PLANNER_MODEL` | `opus` | Model for planning agent |\n| `GAN_GENERATOR_MODEL` | `opus` | Model for generator agent |\n| `GAN_EVALUATOR_MODEL` | `opus` | Model for evaluator agent |\n| `GAN_EVAL_CRITERIA` | `design,originality,craft,functionality` | Comma-separated criteria |\n| `GAN_DEV_SERVER_PORT` | `3000` | Port for the live app |\n| `GAN_DEV_SERVER_CMD` | `npm run dev` | Command to start dev server |\n| `GAN_PROJECT_DIR` | `.` | Project working directory |\n| `GAN_SKIP_PLANNER` | `false` | Skip planner, use spec directly |\n| `GAN_EVAL_MODE` | `playwright` | `playwright`, `screenshot`, or `code-only` |\n\n### Evaluation Modes\n\n| Mode | Tools | Best For |\n|------|-------|----------|\n| `playwright` | Browser MCP + live interaction | Full-stack apps with UI |\n| `screenshot` | Screenshot + visual analysis | Static sites, design-only |\n| `code-only` | Tests + linting + build | APIs, libraries, CLI tools |\n\n## Anti-Patterns\n\n1. **Evaluator too lenient** — If the evaluator passes everything on iteration 1, your rubric is too generous. Tighten scoring criteria and add explicit penalties for common AI patterns.\n\n2. **Generator ignoring feedback** — Ensure feedback is passed as a file, not inline. The generator should read `feedback-NNN.md` at the start of each iteration.\n\n3. **Infinite loops** — Always set `GAN_MAX_ITERATIONS`. If the generator can't improve past a score plateau after 3 iterations, stop and flag for human review.\n\n4. **Evaluator testing superficially** — The evaluator must use Playwright to **interact** with the live app, not just screenshot it. Click buttons, fill forms, test error states.\n\n5. **Evaluator praising its own fixes** — Never let the evaluator suggest fixes and then evaluate those fixes. The evaluator only critiques; the generator fixes.\n\n6. **Context exhaustion** — For long sessions, use Claude Agent SDK's automatic compaction or reset context between major phases.\n\n## Results: What to Expect\n\nBased on Anthropic's published results:\n\n| Metric | Solo Agent | GAN Harness | Improvement |\n|--------|-----------|-------------|-------------|\n| Time | 20 min | 4-6 hours | 12-18x longer |\n| Cost | $9 | $125-200 | 14-22x more |\n| Quality | Barely functional | Production-ready | Phase change |\n| Core features | Broken | All working | N\u002FA |\n| Design | Generic AI slop | Distinctive, polished | N\u002FA |\n\n**The tradeoff is clear:** ~20x more time and cost for a qualitative leap in output quality. This is for projects where quality matters.\n\n## References\n\n- [Anthropic: Harness Design for Long-Running Apps](https:\u002F\u002Fwww.anthropic.com\u002Fengineering\u002Fharness-design-long-running-apps) — Original paper by Prithvi Rajasekaran\n- [Epsilla: The GAN-Style Agent Loop](https:\u002F\u002Fwww.epsilla.com\u002Fblogs\u002Fanthropic-harness-engineering-multi-agent-gan-architecture) — Architecture deconstruction\n- [Martin Fowler: Harness Engineering](https:\u002F\u002Fmartinfowler.com\u002Farticles\u002Fexploring-gen-ai\u002Fharness-engineering.html) — Broader industry context\n- [OpenAI: Harness Engineering](https:\u002F\u002Fopenai.com\u002Findex\u002Fharness-engineering\u002F) — OpenAI's parallel work\n","# GAN-Style Harness Skill\n\n> Inspired by [Anthropic's Harness Design for Long-Running Application Development](https:\u002F\u002Fwww.anthropic.com\u002Fengineering\u002Fharness-design-long-running-apps) (March 24, 2026)\n\nA multi-agent harness that separates **generation** from **evaluation**, creating an adversarial feedback loop that drives quality far beyond what a single agent can achieve.\n\n## Core Insight\n\n> When asked to evaluate their own work, agents are pathological optimists — they praise mediocre output and talk themselves out of legitimate issues. But engineering a **separate evaluator** to be ruthlessly strict is far more tractable than teaching a generator to self-critique.\n\nThis is the same dynamic as GANs (Generative Adversarial Networks): the Generator produces, the Evaluator critiques, and that feedback drives the next iteration.\n\n## When to Use\n\n- Building complete applications from a one-line prompt\n- Frontend design tasks requiring high visual quality\n- Full-stack projects that need working features, not just code\n- Any task where \"AI slop\" aesthetics are unacceptable\n- Projects where you want to invest $50-200 for production-quality output\n\n## When NOT to Use\n\n- Quick single-file fixes (use standard `claude -p`)\n- Tasks with tight budget constraints (\u003C$10)\n- Simple refactoring (use de-sloppify pattern instead)\n- Tasks that are already well-specified with tests (use TDD workflow)\n\n## Architecture\n\n```\n                    ┌─────────────┐\n                    │   PLANNER   │\n                    │  (Opus 4.6) │\n                    └──────┬──────┘\n                           │ Product Spec\n                           │ (features, sprints, design direction)\n                           ▼\n              ┌────────────────────────┐\n              │                        │\n              │   GENERATOR-EVALUATOR  │\n              │      FEEDBACK LOOP     │\n              │                        │\n              │  ┌──────────┐          │\n              │  │GENERATOR │--build-->│──┐\n              │  │(Opus 4.6)│          │  │\n              │  └────▲─────┘          │  │\n              │       │                │  │ live app\n              │    feedback             │  │\n              │       │                │  │\n              │  ┌────┴─────┐          │  │\n              │  │EVALUATOR │\u003C-test----│──┘\n              │  │(Opus 4.6)│          │\n              │  │+Playwright│         │\n              │  └──────────┘          │\n              │                        │\n              │   5-15 iterations      │\n              └────────────────────────┘\n```\n\n## The Three Agents\n\n### 1. Planner Agent\n\n**Role:** Product manager — expands a brief prompt into a full product specification.\n\n**Key behaviors:**\n- Takes a one-line prompt and produces a 16-feature, multi-sprint specification\n- Defines user stories, technical requirements, and visual design direction\n- Is deliberately **ambitious** — conservative planning leads to underwhelming results\n- Produces evaluation criteria that the Evaluator will use later\n\n**Model:** Opus 4.6 (needs deep reasoning for spec expansion)\n\n### 2. Generator Agent\n\n**Role:** Developer — implements features according to the spec.\n\n**Key behaviors:**\n- Works in structured sprints (or continuous mode with newer models)\n- Negotiates a \"sprint contract\" with the Evaluator before writing code\n- Uses full-stack tooling: React, FastAPI\u002FExpress, databases, CSS\n- Manages git for version control between iterations\n- Reads Evaluator feedback and incorporates it in next iteration\n\n**Model:** Opus 4.6 (needs strong coding capability)\n\n### 3. Evaluator Agent\n\n**Role:** QA engineer — tests the live running application, not just code.\n\n**Key behaviors:**\n- Uses **Playwright MCP** to interact with the live application\n- Clicks through features, fills forms, tests API endpoints\n- Scores against four criteria (configurable):\n  1. **Design Quality** — Does it feel like a coherent whole?\n  2. **Originality** — Custom decisions vs. template\u002FAI patterns?\n  3. **Craft** — Typography, spacing, animations, micro-interactions?\n  4. **Functionality** — Do all features actually work?\n- Returns structured feedback with scores and specific issues\n- Is engineered to be **ruthlessly strict** — never praises mediocre work\n\n**Model:** Opus 4.6 (needs strong judgment + tool use)\n\n## Evaluation Criteria\n\nThe default four criteria, each scored 1-10:\n\n```markdown\n## Evaluation Rubric\n\n### Design Quality (weight: 0.3)\n- 1-3: Generic, template-like, \"AI slop\" aesthetics\n- 4-6: Competent but unremarkable, follows conventions\n- 7-8: Distinctive, cohesive visual identity\n- 9-10: Could pass for a professional designer's work\n\n### Originality (weight: 0.2)\n- 1-3: Default colors, stock layouts, no personality\n- 4-6: Some custom choices, mostly standard patterns\n- 7-8: Clear creative vision, unique approach\n- 9-10: Surprising, delightful, genuinely novel\n\n### Craft (weight: 0.3)\n- 1-3: Broken layouts, missing states, no animations\n- 4-6: Works but feels rough, inconsistent spacing\n- 7-8: Polished, smooth transitions, responsive\n- 9-10: Pixel-perfect, delightful micro-interactions\n\n### Functionality (weight: 0.2)\n- 1-3: Core features broken or missing\n- 4-6: Happy path works, edge cases fail\n- 7-8: All features work, good error handling\n- 9-10: Bulletproof, handles every edge case\n```\n\n### Scoring\n\n- **Weighted score** = sum of (criterion_score * weight)\n- **Pass threshold** = 7.0 (configurable)\n- **Max iterations** = 15 (configurable, typically 5-15 sufficient)\n\n## Usage\n\n### Via Command\n\n```bash\n# Full three-agent harness\n\u002Fproject:gan-build \"Build a project management app with Kanban boards, team collaboration, and dark mode\"\n\n# With custom config\n\u002Fproject:gan-build \"Build a recipe sharing platform\" --max-iterations 10 --pass-threshold 7.5\n\n# Frontend design mode (generator + evaluator only, no planner)\n\u002Fproject:gan-design \"Create a landing page for a crypto portfolio tracker\"\n```\n\n### Via Shell Script\n\n```bash\n# Basic usage\n.\u002Fscripts\u002Fgan-harness.sh \"Build a music streaming dashboard\"\n\n# With options\nGAN_MAX_ITERATIONS=10 \\\nGAN_PASS_THRESHOLD=7.5 \\\nGAN_EVAL_CRITERIA=\"functionality,performance,security\" \\\n.\u002Fscripts\u002Fgan-harness.sh \"Build a REST API for task management\"\n```\n\n### Via Claude Code (Manual)\n\n```bash\n# Step 1: Plan\nclaude -p --model opus \"You are a Product Planner. Read PLANNER_PROMPT.md. Expand this brief into a full product spec: 'Build a Kanban board app'. Write spec to spec.md\"\n\n# Step 2: Generate (iteration 1)\nclaude -p --model opus \"You are a Generator. Read spec.md. Implement Sprint 1. Start the dev server on port 3000.\"\n\n# Step 3: Evaluate (iteration 1)\nclaude -p --model opus --allowedTools \"Read,Bash,mcp__playwright__*\" \"You are an Evaluator. Read EVALUATOR_PROMPT.md. Test the live app at http:\u002F\u002Flocalhost:3000. Score against the rubric. Write feedback to feedback-001.md\"\n\n# Step 4: Generate (iteration 2 — reads feedback)\nclaude -p --model opus \"You are a Generator. Read spec.md and feedback-001.md. Address all issues. Improve the scores.\"\n\n# Repeat steps 3-4 until pass threshold met\n```\n\n## Evolution Across Model Capabilities\n\nThe harness should simplify as models improve. Following Anthropic's evolution:\n\n### Stage 1 — Weaker Models (Sonnet-class)\n- Full sprint decomposition required\n- Context resets between sprints (avoid context anxiety)\n- 2-agent minimum: Initializer + Coding Agent\n- Heavy scaffolding compensates for model limitations\n\n### Stage 2 — Capable Models (Opus 4.5-class)\n- Full 3-agent harness: Planner + Generator + Evaluator\n- Sprint contracts before each implementation phase\n- 10-sprint decomposition for complex apps\n- Context resets still useful but less critical\n\n### Stage 3 — Frontier Models (Opus 4.6-class)\n- Simplified harness: single planning pass, continuous generation\n- Evaluation reduced to single end-pass (model is smarter)\n- No sprint structure needed\n- Automatic compaction handles context growth\n\n> **Key principle:** Every harness component encodes an assumption about what the model can't do alone. When models improve, re-test those assumptions. Strip away what's no longer needed.\n\n## Configuration\n\n### Environment Variables\n\n| Variable | Default | Description |\n|----------|---------|-------------|\n| `GAN_MAX_ITERATIONS` | `15` | Maximum generator-evaluator cycles |\n| `GAN_PASS_THRESHOLD` | `7.0` | Weighted score to pass (1-10) |\n| `GAN_PLANNER_MODEL` | `opus` | Model for planning agent |\n| `GAN_GENERATOR_MODEL` | `opus` | Model for generator agent |\n| `GAN_EVALUATOR_MODEL` | `opus` | Model for evaluator agent |\n| `GAN_EVAL_CRITERIA` | `design,originality,craft,functionality` | Comma-separated criteria |\n| `GAN_DEV_SERVER_PORT` | `3000` | Port for the live app |\n| `GAN_DEV_SERVER_CMD` | `npm run dev` | Command to start dev server |\n| `GAN_PROJECT_DIR` | `.` | Project working directory |\n| `GAN_SKIP_PLANNER` | `false` | Skip planner, use spec directly |\n| `GAN_EVAL_MODE` | `playwright` | `playwright`, `screenshot`, or `code-only` |\n\n### Evaluation Modes\n\n| Mode | Tools | Best For |\n|------|-------|----------|\n| `playwright` | Browser MCP + live interaction | Full-stack apps with UI |\n| `screenshot` | Screenshot + visual analysis | Static sites, design-only |\n| `code-only` | Tests + linting + build | APIs, libraries, CLI tools |\n\n## Anti-Patterns\n\n1. **Evaluator too lenient** — If the evaluator passes everything on iteration 1, your rubric is too generous. Tighten scoring criteria and add explicit penalties for common AI patterns.\n\n2. **Generator ignoring feedback** — Ensure feedback is passed as a file, not inline. The generator should read `feedback-NNN.md` at the start of each iteration.\n\n3. **Infinite loops** — Always set `GAN_MAX_ITERATIONS`. If the generator can't improve past a score plateau after 3 iterations, stop and flag for human review.\n\n4. **Evaluator testing superficially** — The evaluator must use Playwright to **interact** with the live app, not just screenshot it. Click buttons, fill forms, test error states.\n\n5. **Evaluator praising its own fixes** — Never let the evaluator suggest fixes and then evaluate those fixes. The evaluator only critiques; the generator fixes.\n\n6. **Context exhaustion** — For long sessions, use Claude Agent SDK's automatic compaction or reset context between major phases.\n\n## Results: What to Expect\n\nBased on Anthropic's published results:\n\n| Metric | Solo Agent | GAN Harness | Improvement |\n|--------|-----------|-------------|-------------|\n| Time | 20 min | 4-6 hours | 12-18x longer |\n| Cost | $9 | $125-200 | 14-22x more |\n| Quality | Barely functional | Production-ready | Phase change |\n| Core features | Broken | All working | N\u002FA |\n| Design | Generic AI slop | Distinctive, polished | N\u002FA |\n\n**The tradeoff is clear:** ~20x more time and cost for a qualitative leap in output quality. This is for projects where quality matters.\n\n## References\n\n- [Anthropic: Harness Design for Long-Running Apps](https:\u002F\u002Fwww.anthropic.com\u002Fengineering\u002Fharness-design-long-running-apps) — Original paper by Prithvi Rajasekaran\n- [Epsilla: The GAN-Style Agent Loop](https:\u002F\u002Fwww.epsilla.com\u002Fblogs\u002Fanthropic-harness-engineering-multi-agent-gan-architecture) — Architecture deconstruction\n- [Martin Fowler: Harness Engineering](https:\u002F\u002Fmartinfowler.com\u002Farticles\u002Fexploring-gen-ai\u002Fharness-engineering.html) — Broader industry context\n- [OpenAI: Harness Engineering](https:\u002F\u002Fopenai.com\u002Findex\u002Fharness-engineering\u002F) — OpenAI's parallel work\n",{"name":5,"tools":24,"origin":25,"description":7},"Read, Write, Edit, Bash, Grep, Glob, Task","ECC-community","master","d6ca01da0f24a3b0e26e4a8b7b66cbd25d429f5d","skills\u002Fgan-style-harness",0,"https:\u002F\u002Fgithub.com\u002FSriramGandhiS\u002FSkills","https:\u002F\u002Fgithub.com\u002FSriramGandhiS","https:\u002F\u002Fsriram.website\u002F","Curious guy who keeps tinkering until something works","2026-05-20 16:21:34",{"source":8,"skillName":5,"commands":36,"prompt":46,"zipAvailable":47},[37,40,43],{"tool":38,"command":39},"npx","npx skills add SriramGandhiS\u002FSkills --skill gan-style-harness",{"tool":41,"command":42},"bunx","bunx skills add SriramGandhiS\u002FSkills --skill gan-style-harness",{"tool":44,"command":45},"pnpm","pnpm dlx skills add SriramGandhiS\u002FSkills --skill gan-style-harness","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FSriramGandhiS\u002FSkills\u002Fblob\u002Fmaster\u002Fskills\u002Fgan-style-harness\u002FSKILL.md",true,[],[50,57,62,68,75,82],{"skillNo":51,"name":52,"nameZh":6,"description":53,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":54,"categories":55,"keywords":56,"publishedAt":19},"01M732609EPJWV13PVTX3GT4TM","webgpu-threejs-tsl","Comprehensive guide for developing WebGPU-enabled Three.js applications using TSL (Three.js Shading Language). Covers WebGPU renderer setup, TSL syntax and node materials, compute shaders, post-processing effects, and WGSL integration. Use this skill when working with Three.js WebGPU, TSL shaders, node materials, or GPU compute in Three.js.","https:\u002F\u002Fgithub.com\u002FSriramGandhiS\u002FSkills\u002Fblob\u002Fmaster\u002Fwebgpu-threejs-tsl\u002Fskills\u002Fwebgpu-threejs-tsl\u002FSKILL.md",[],[],{"skillNo":58,"name":52,"nameZh":6,"description":53,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":59,"categories":60,"keywords":61,"publishedAt":19},"01M7325XM47AGWB6K6C4VXKE6B","https:\u002F\u002Fgithub.com\u002FSriramGandhiS\u002FSkills\u002Fblob\u002Fmaster\u002Fwebgpu-threejs-tsl\u002FSKILL.md",[],[],{"skillNo":63,"name":64,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":65,"categories":66,"keywords":67,"publishedAt":19},"01M7325WP6DES69JKC840VMFEP","verification-loop","https:\u002F\u002Fgithub.com\u002FSriramGandhiS\u002FSkills\u002Fblob\u002Fmaster\u002Fweb-assets\u002Fdocs\u002Fzh-TW\u002Fskills\u002Fverification-loop\u002FSKILL.md",[],[],{"skillNo":69,"name":70,"nameZh":6,"description":71,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":72,"categories":73,"keywords":74,"publishedAt":19},"01M7325TFQF1DZTQ6J3WGQQ1VR","tdd-workflow","Use this skill when writing new features, fixing bugs, or refactoring code. Enforces test-driven development with 80%+ coverage including unit, integration, and E2E tests.","https:\u002F\u002Fgithub.com\u002FSriramGandhiS\u002FSkills\u002Fblob\u002Fmaster\u002Fweb-assets\u002Fdocs\u002Fzh-TW\u002Fskills\u002Ftdd-workflow\u002FSKILL.md",[],[],{"skillNo":76,"name":77,"nameZh":6,"description":78,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":79,"categories":80,"keywords":81,"publishedAt":19},"01M7325RGV9AY9PK92D2V8KKPW","strategic-compact","Suggests manual context compaction at logical intervals to preserve context through task phases rather than arbitrary auto-compaction.","https:\u002F\u002Fgithub.com\u002FSriramGandhiS\u002FSkills\u002Fblob\u002Fmaster\u002Fweb-assets\u002Fdocs\u002Fzh-TW\u002Fskills\u002Fstrategic-compact\u002FSKILL.md",[],[],{"skillNo":83,"name":84,"nameZh":6,"description":85,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":86,"categories":87,"keywords":88,"publishedAt":19},"01M7325MZC28SE86E7NMCNEFS4","security-review","Use this skill when adding authentication, handling user input, working with secrets, creating API endpoints, or implementing payment\u002Fsensitive features. Provides comprehensive security checklist and patterns.","https:\u002F\u002Fgithub.com\u002FSriramGandhiS\u002FSkills\u002Fblob\u002Fmaster\u002Fweb-assets\u002Fdocs\u002Fzh-TW\u002Fskills\u002Fsecurity-review\u002FSKILL.md",[],[]]